Short-Term Wind Power Interval Prediction Based on GD-LSTM and Bootstrap Techniques

Wind power forecasting plays a significant role in the stability, security and reliability of power system. Compared with point forecast, interval forecast is able to quantify the uncertainty effectively. In this paper, a novel interval forecast method based on modified Gaussian-Dropout long short-term memory (GD-LSTM) network and block bootstrap is proposed to predict short-term wind power. In this model, prediction uncertainty is concerned to come from model parameters and data noise. Bootstrap is used to generate multiple subsets firstly, and then LSTM models are trained by these subsets. The test dataset prediction intervals (PIs) are generated with confidence levels ranging from 85% to 95%. The results demonstrate that the PIs generated by proposed method can always reach the targets, the coverage probability generated by block bootstrap may lower than other methods in sometimes, while the comprehensive scores are always the best.